RTM — Recursive Transformer Model

Self-Referential Inference • Persistent Memory • KV-Slotting • ARC-AGI Performance

Author: Dr. Josef Kurk Edwards (drQedwards)
Status: Research Release
License: MIT

---

Recursive Transformer Model (RTM)

Overview

The Recursive Transformer Model (RTM) extends standard transformers with persistent memory, temporal decay, consensus validation, and contradiction detection, enabling stateful reasoning across inference sessions. This directly addresses the problem of nostalgic incorrectness—a model’s tendency to retain outdated or contradicted beliefs.

RTM draws from Dr. Josef “Q.” Edwards’ 2025 research on Persistent Memory Logic Loops (PMLL) and the Enhanced Reconsideration System (ERS). RTM is complementary to the Tiny Recursion Model (TRM), and a hybrid TRM–RTM architecture is possible.

---

Features

✅ Persistent Memory Store

RTM stores past outputs in vectorized memory blocks with:

• Timestamped entries
• Confidence scores
• Embedding-based similarity lookup

✅ Temporal Decay

Older or unsupported memories naturally lose confidence, reducing stale knowledge.

✅ Consensus Strengthening

Related memories reinforce one another, improving model reliability.

✅ Contradiction Detection

Conflicting memories penalize each other’s confidence, reducing knowledge drift.

✅ PMLL (Persistent Memory Logic Loop) Scaffold

The repository includes a placeholder for:

• Lattice-based tensor routing
• Memory hashing
• Commitment and re-evaluation cycles

---

Installation

git clone <your repo URL>
cd recursive_transformer
pip install -r requirements.txt

Requirements:

• Python 3.10+
• numpy
• scikit-learn
• dateutil

---

Basic Usage

Example: using a dummy echo model

from recursive_transformer import RecursiveTransformerModel

def echo_model(prompt: str):
    return prompt

model = RecursiveTransformerModel(base_model=echo_model)

result, confidence = model.generate("What is memory?")
print(result, confidence)

---

File Structure

recursive_transformer/
│
├── recursive_transformer/
│   ├── memory.py
│   ├── model.py
│   ├── consensus.py
│   ├── contradiction.py
│   ├── decay.py
│   ├── pmll.py
│   └── __init__.py
│
└── examples/
    └── simple_example.py

---

Comparison: TRM vs RTM vs Hybrid

Feature	TRM	RTM	Hybrid
Primary skill	Recursive reasoning	Persistent state	Both
Memory	Stateless	Persistent & decaying	Persistent + recursive
Best for	ARC puzzles	Knowledge systems	AGI-level reasoning
Parameter size	~7M	Model-dependent	Model-dependent

---

Citation

If you use RTM in research:

Edwards, J. (2025). The Recursive Transformer Model: Architecture, Theory, and Implementation with Persistent Memory Logic Loops.

---

License

MIT License

RTM is an experimental transformer architecture that fuses:

• Recursive forward passes (self-calling inference path)
• PMLL — Persistent Memory Logic Loop
• KV-Slotting for long-context efficiency
• ARC-AGI benchmark evaluation modules

RTM explores the hypothesis:

A transformer equipped with recursive inference and persistent memory can solve abstraction tasks that defeat standard models.

---

Features

Recursive Core

The model calls itself under a continuation policy, enabling multistep reasoning chains.

PMLL Integration

Hash-chained semantic memory with Race/Trace retrieval and compressed persistence.

KV-Slotting

Selective KV-cache rehydration for long-horizon inference without memory blowup.

ARC-AGI Evaluation

Includes adapters, evaluation loops, scoring, and ablation hooks.

---

Installation

git clone https://github.com/drQedwards/RTM.git
cd RTM
unzip recursive_transformer.zip -d rtm_src
pip install -r requirements.txt

---

ARC-AGI Benchmarking

The benchmarks/arc/ module provides:

• Task loading
• Model ↔ ARC adapters
• Scoring
• Recursion-depth analysis
• Memory hit-rate analysis

Run:

python benchmarks/arc/run_arc.py --model rtm --config configs/arc/rtm_default.yaml

---

Roadmap

• Automated ARC reporting
• Memory graph visualizer
• Docker reproducibility
• TechRxiv whitepaper

---

License

MIT License. See LICENSE for details.